UK What MGX is
NVIDIA MGX is a modular reference architecture that server manufacturers use to design accelerated computers. NVIDIA announced it at Computex in May 2023. Rather than each manufacturer engineering a new GPU server from scratch for every processor and GPU generation, MGX gives them a common set of validated building blocks: chassis layouts, GPU and CPU board positions, power delivery, cooling and networking slots.
Each manufacturer then combines those blocks into its own products and adds its own motherboard, firmware, management and support. NVIDIA said at launch that more than 100 configurations were possible, and that the approach could cut development cost by up to three-quarters and bring a new design to market in around six months. Those are NVIDIA’s projections for manufacturers, not figures a buyer will see on a quote.
So MGX is not a product you buy from NVIDIA. It is the design specification behind a family of servers that different manufacturers sell under their own names.
What MGX standardises
MGX fixes the parts of a GPU server that are hardest to get right, while leaving room for each maker to differentiate.
-
//
Chassis and mechanical layout. Standard 1U, 2U and 4U designs and rack-scale trays, with defined spacing for double-width GPUs and front-to-back airflow.
-
//
Host processor. Interchangeable CPU sections for Intel Xeon, AMD EPYC and NVIDIA’s Arm-based Grace, so one chassis design can carry more than one processor family.
-
//
Accelerators. Slots for PCIe GPUs such as the H200 NVL, L40S and RTX PRO 6000 Blackwell Server Edition, or Grace Hopper and Grace Blackwell superchips on dedicated boards.
-
//
Networking. Positions for ConnectX network adapters and BlueField DPUs, which take networking and security work off the host CPU, commonly in OCP 3.0 slots, for linking servers into a cluster.
-
//
Power and cooling. Defined power supplies and both air-cooled and liquid-cooled options. NVIDIA has also contributed MGX rack designs to the Open Compute Project (OCP).
MGX vs HGX vs DGX
The three names sit at different levels, and they are easy to confuse because all three end up as GPU servers.
| MGX | HGX | DGX | |
|---|---|---|---|
| What it is | Modular reference architecture for many server designs | NVIDIA GPU baseboard with SXM GPUs and NVSwitch | Complete NVIDIA-branded AI system |
| Who builds the server | Server manufacturers, to their own design | Server manufacturers, around NVIDIA’s board | NVIDIA |
| Typical GPUs | PCIe cards or Grace superchips | 4 or 8 SXM GPUs (H100, H200, B200) | 8 SXM GPUs or superchips |
| GPU-to-GPU link | NVLink bridges, or NVLink in superchip racks | NVSwitch, all-to-all | NVSwitch, all-to-all |
| Flexibility | Widest: CPU, GPU count, size and cooling vary | Fixed GPU board, flexible host | Fixed configuration |
| Typical use | Inference, mixed AI and graphics, enterprise and edge racks | Large-scale training and inference | Complete AI clusters from one supplier |
In practice, an MGX server with four PCIe GPUs suits inference and mixed workloads in a standard air-cooled rack. HGX and DGX systems suit training jobs that need every GPU linked to every other at full speed.
Where MGX servers fit
Because MGX supports PCIe GPUs in ordinary 19-inch racks, many MGX designs target enterprise server rooms rather than hyperscale halls. A 2U system with two Intel Xeon 6 processors and four double-width GPUs, such as the SKY-622G4 MGX GPU server, can run a large language model, a retrieval pipeline and several vision models side by side. It takes NVLink bridges between its cards and front-loading NVMe storage.
At the other end of the range, NVIDIA’s liquid-cooled GB200 NVL72 racks also build on MGX rack designs. Our rack mount servers range covers GPU servers alongside conventional rack systems, and our data centre GPUs are the cards these servers take.
What to check before choosing an MGX system
MGX narrows the engineering risk, but it does not make every part interchangeable. Each finished server supports a specific list of GPUs, processors and adapters.
-
//
GPU list and count. Confirm the exact GPUs the system is validated for, how many double-width cards fit and whether NVLink bridges are supported at that slot spacing.
-
//
Power and cooling. Four high-power GPUs can take a 2U server to around 3 kW or more. Check the power supplies, the rack’s power feed and whether the room can remove that heat by air.
-
//
Depth and weight. GPU servers are often 800–900 mm deep, so check the rack depth and rails.
-
//
Management. Look for a baseboard management controller with Redfish and IPMI support for remote monitoring and firmware updates.
-
//
Networking. Plan the cluster network early: the NIC or DPU, its port speed and the switch it connects to.
Quick answers on NVIDIA MGX
These are the questions buyers most often ask when MGX appears on a server datasheet.
NVIDIA MGX: common questions
Can I buy an MGX server from NVIDIA?
Not directly. MGX is a design specification. NVIDIA’s own systems are sold as DGX; MGX servers come from server manufacturers and the integrators who build with them.
Is MGX the same as HGX?
No. HGX is an NVIDIA GPU baseboard carrying SXM GPUs linked by NVSwitch. MGX is a wider modular design framework, mostly built around PCIe GPUs and Grace superchips.
Does an MGX server need liquid cooling?
No. Many MGX servers with PCIe GPUs are air-cooled. Liquid cooling applies to the densest designs, such as rack-scale Grace Blackwell systems.
Which processors work in MGX servers?
x86 and Arm. MGX designs support Intel Xeon and AMD EPYC as well as NVIDIA’s Grace CPU, depending on the specific server.
Does MGX support NVLink?
Yes, in two ways. PCIe GPUs in MGX servers can be paired with NVLink bridges, and superchip-based MGX racks use NVLink across the whole rack.
We build MGX and other GPU servers with the GPUs, memory, storage and network adapters fitted and the operating system and drivers installed as standard, and we can help you match the GPU count to your power and cooling.